Skip to content- Sources: primary
- Summary: Meta engineers say about three quarters of new code is AI-written, and they replaced per-silo bots with one agent per rotation that has its own identity, persistent team memory and a morning digest. At level one it investigates, at level two it drafts fixes that land only through normal review and CI, and at level three it handles low-criticality work such as noisy alerts and duplicates. Memory entries are scored on whether they helped and are promoted or retired, and the team has not yet published a measured on-call reduction.
- Why it matters: It gives a concrete pattern for granting agent autonomy per action, based on mistake cost and how easily tests or review catch errors.